BiasTrust: Teaching Biased Users About Controversial
Topics
†
†
V.G.Vinod Vydiswaran
ChengXiang Zhai
†
Dan Roth
‡
Peter Pirolli
†
University of Illinois, Urbana, IL
Palo Alto Research Center, Palo Alto, CA
‡
†
‡
{vgvinodv,czhai,danr}@illinois.edu, [email protected]
ABSTRACT
ible sources have made significant errors in stating facts or
helped spread rumors; possibly due to their own biases. This
indicates a growing challenge for users to gather an unbiased
opinion about topics of interest over the Internet and also
provides a strong motivation to study it.
Consider a scenario where an Internet surfer wants to
learn about a recent controversy. To give an example, consider that Alice wants to know if chocolate and flavored
milk provided in schools is a healthy food choice for her
kids. Depending on the keywords she chooses to search
about the topic, she might see news articles about a recent ban on chocolate milk in certain schools, or learn about
health benefits of milk in growing kids. She might find results from news media organizations reporting on the ban, or
activist groups actively encouraging drinking milk, or even
concerned parents posting questions and responding to them
via community-driven question answering services.
Is Alice equally likely to read these articles, and read them
in the order presented? The answers appear to be in the
negative, but we wanted to understand what factors impact
these decisions. Specifically, we wanted to study the factors
that influenced (i) the documents users read, (ii) the extent
of learning, and (iii) the perceived credibility of a source.
We believe this is one of the first works that studies these
aspects, especially when learning about controversial topics.
We designed and conducted a user study to understand
how to present evidence documents to enable users to learn
about the controversy in an unbiased fashion and help them
overcome any prior bias they might have about the topic.
First, we try to understand the user’s position and beliefs
about the controversy. Then, we study if the display of contrasting viewpoints to evidence documents and source expertise ratings influence user behavior. We find that users tend
not to seek contrasting viewpoints, at least in a limited-time
learning scenario, but that explicitly presenting contrasting
evidence helps them get a balanced understanding of the
topic. Furthermore, we observe that explicit knowledge of
the source credibility or expertise, and the context in which
the evidence was provided, not only affects what users read,
but also how credible they perceive the documents to be.
These insights help us optimize the presentation of credible
evidence documents to teach biased users about controversial topics in the most effective way.
In this short paper, we explain the design of the user study
and the interface variants in detail. However, due to space
constraints, we restrict our analysis to factors that help users
get a balanced understanding of the topic. Additional findings are presented in Vydiswaran et al. [9].
Deciding whether a claim is true or false often requires understanding the evidence supporting and contradicting the
claim. However, when learning about a controversial claim,
human biases and viewpoints may affect which evidence documents are considered “trustworthy” or credible. It is important to overcome this bias and know both viewpoints to
get a balanced perspective. In this paper, we study various
factors that affect learning about the truthfulness of controversial claims. We designed a user study to understand
the impact of these factors. Specifically, we studied the impact of presenting evidence with contrasting viewpoints and
source expertise rating on how users accessed the evidence
documents. This would help us optimize how to teach users
about controversial topics in the most effective way, and
to design better claim verification systems. We find that
users do not seek contrasting viewpoints by themselves, but
explicitly presenting contrasting evidence helps them get a
well-rounded understanding of the topic. Furthermore, explicit knowledge of the source credibility and the context
not only affects what users read, but also how credible they
perceive the document to be.
Categories and Subject Descriptors
H.1.2 [Information Systems]: Models and Principles—
Human information processing, Human factors
Keywords
Claim verification, Information credibility, User study
1.
INTRODUCTION
The Web today is a hodgepodge of well-curated, edited
content and freelance, unmoderated content. Typically, wellstructured, formatted, and edited content is considered more
trustworthy and credible, while information that appears in
forums and message boards is considered not as trustworthy. But history is replete with many instances where cred-
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
CIKM’12, October 29–November 2, 2012, Maui, HI, USA.
Copyright 2012 ACM 978-1-4503-1156-4/12/10 ...$15.00.
1905
2.
RELATED WORK
points about a claim to help gauge the trustworthiness of the
claim; (ii) the knowledge of source expertise and credibility,
and how that affects the decision on what evidence is read;
and (iii) the order in which the documents are presented,
and if that affects the overall understanding of the topic.
Other factors, such as summarizing documents and assigning an overall truth value to a claim may also be relevant
in designing a claim verification system. However, we chose
not to study these factors, but instead provide access to the
relevant evidence to allow users verify claims by themselves.
Further, the study focused on controversial claims instead
of factual claims. Controversial claims have many genuine
evidence documents supporting and opposing the claims, so
by focusing on such claims, we could study how preferencebased factors affect the learning process.
The user study was designed as a learning task, where
users were asked to learn as much as possible about a topic
within a stipulated time. The study was conducted in three
online stages, viz. (i) Pre-study survey questionnaire, (ii) the
Study phase, and (iii) Post-study questionnaire.
Stage 1: Pre-study questionnaire: In the pre-study
survey, subjects were asked questions that helped us gauge
their (lack of) knowledge and bias towards/against issues
relevant to the topic being studied. Subjects answered the
questions on a four-point Likert scale. For the knowledge
questions, the response scale ranged from (i) ‘insignificant’
to (iv) ‘very significant’, while for bias questions, it ranged
from (i) ‘strongly against’ to (iv) ‘strongly in favor of’ the
issue. Subjects could also respond to a question with an ‘I
don’t know’ answer. This design of using knowledge and
bias related questions and limiting the nature of allowed
responses was to encourage subjects to think about their
position on many sub-topics related to the overall issue.
Stage 2: Study phase: Once the responses to the presurvey questionnaire were recorded, subjects were directed
to one of the interface variants (described in detail in Sec. 4).
In each interface variant, subjects had some contextual information about the passages. For instance, subjects were
shown the source of the passage, its sub-topic, and whether
the passage was in favor or against the sub-topic. For each
passage subjects read, they were asked to answer two questions about the passage, viz. whether they agreed with what
was being said in the passage and if they believed the information was biased with respect to the topic. Subjects were
asked to continue reading until they believed they had read
enough about the topic. Once they decided to quit the study
phase, they were taken to the third and final stage of the
online study.
Stage 3: Post-study questionnaire: After subjects
spent time learning about the topic in the study phase, they
were asked to answer the same set of topic-specific questions
that were posed during the pre-study questionnaire. Subjects responded based on the topics they read about and
how important and relevant they felt the sub-topics were,
after the study. They were also asked to provide feedback
on what interface features helped them in their task, and
suggest additional features that might help them.
Understanding what documents people read is related to
research in many fields. Psychologists have studied the phenomenon of confirmation bias [5, 1], which states that people
tend to favor information that confirms their beliefs – not
only in deciding what to read, but also how they interpret
what is read. Similarly, researchers in political science [7]
found that people process information in biased fashion.
They uncritically accept supporting arguments and argue
against those that are contrary to their beliefs.
Researchers have looked at aggregating information from
multiple sources to answer specific questions. Wu and Marian [10] studied how to collect information from multiple
sources over the Web, specifically to find answers using corroborative evidence from multiple sources. Gallard et al. [2]
looked at collecting information from contrasting viewpoints.
The work presented in this paper looks at the next step of
how to present these documents with contrasting viewpoints
to users, while providing information about credibility of the
sources, to enable users to learn about the topic efficiently.
Researchers have also looked at factors that influence what
information users access and how they process them. This
includes work on building tools to increase transparency and
credibility of Wikipedia articles [4, 6] to help users decide if
the information is credible or not. Ugander et al. [8] studied how to convince people based on the influence of one’s
social network on their actions. Extending it to the information domain, this research suggests that users should be
exposed to multiple viewpoints. Pariser [3] investigated the
notion of filter bubble, where search engines personalize web
search results to show users information similar to what they
have already seen. This not only encourages confirmation
bias, but also excludes contradictory viewpoints. Our study
tries to understand how to overcome these shortcomings by
presenting contrasting, yet credible evidence to users. We
believe our work is the first to study how source expertise
rating, evidence context, and contrasting viewpoints help
users learn about controversial topics.
3.
BIASTRUST: DESIGNING THE STUDY
Understanding which claims to believe and why to believe those claims is important to make an informed decision. This is basically a learning task, where an inquisitive
user tries to learn as much as possible about the claim and
assimilate all evidence in support of or against the claim. It
is an interesting challenge for a retrieval system to not only
retrieve documents relevant to the claim, but also present it
succinctly to help users understand that information quickly.
However, as we pointed out in Sec. 2, previous research by
psychologists and others have shown that users tend to access information that supports their own viewpoints. So, it
is important for an automated claim verification system to
model such human biases and present trustworthy evidence
to overcome this bias, where possible.
We designed a system that would retrieve relevant, trustworthy documents about a topic and provide the user with
an overall, unbiased perspective. We conducted a user study
to learn how to optimize the display of credible information to users. The BiasTrust study investigates the factors
that may influence humans to decide what to read and what
to treat as credible. We focused on three major factors,
viz. (i) the ability to access credible, yet contrasting view-
4. USER INTERFACE VARIANTS
We designed interfaces to study the following factors: (i) explicit display of contrasting evidence, (ii) single document
per page vs. multiple documents, (iii) source expertise rating, and (iv) presentation order of documents.
1906
Figure 1: Single document view with option to look at
contrasting document (UI# {1a, 1b})
Figure 3: Multi-document view that shows five priFigure 2: Single document view that shows contrasting mary documents, along with the corresponding condocument by default (UI# {2a, 2b})
trasting documents (UI# {3, 4a, 4b, 5})
Variant
identifier
UI# 1a
UI# 1b
UI# 2a
UI# 2b
UI# 3
UI# 4a
UI# 4b
UI# 5
#
passages
1
1
2
2
10
10
10
10
contrast
view
No
No
Yes
Yes
Yes
Yes
Yes
Yes
topics
sorted
No
No
No
No
No
Yes
Yes
Yes
source rating
show
scheme
Yes
Bimodal
Yes
Uniform
Yes
Bimodal
Yes
Uniform
Yes
Bimodal
Yes
Bimodal
Yes
Uniform
No
—
3. Source expertise rating: Another factor in deciding
what to read is whether users believe the document comes
from a credible source. To test this hypothesis, we decided to
control the expertise ratings in two ways. In one UI variant
(UI# 5), the expertise ratings were hidden, and subjects did
not know if the source was credible or not.
The second way we controlled the expertise rating was to
show two different rating schemes. In one scheme, all sources
got a rating of either 1-star or 3-star. We call this the bimodal rating scheme, since the ratings appear to be drawn
from a bimodal distribution. In the second scheme, sources
were assigned a random rating drawn from a uniform distribution while ensuring that the rating was not the same as
the one under the bimodal scheme. We call this the uniform
rating scheme. Sources were assigned this rating in a static
fashion; that is all participants that were shown one rating
scheme, consistently saw the same expertise rating for the
same source. The UI variants {1a, 2a, 3, 4a} followed the
bimodal rating scheme, while the three remaining variants
{1b, 2b, 4b} followed the uniform rating scheme.
4. Document presentation order: The final hypothesis we wanted to test was whether grouping documents together by sub-topic significantly helped the learning task
and affected the selection of documents read. We focused
on testing this hypothesis only in the case when multiple
documents are shown. We had two configuration settings:
(i) passages are shown in relevance order, and hence appear
to come in a random sequence of sub-topics (in UI variants
{1a, 1b, 2a, 2b, 3}); and (ii) passages are shown in topicsorted order, in which topics and passages from a topic are
both ordered by relevance (in UI variants {4a, 4b, 5}).
Table 1: Parameter settings for interface variants
1. Explicit display of contrasting evidence: We believe that one of the major factors in learning about a controversial topic in an unbiased fashion is the exposure to
alternate viewpoints. To verify if this conjecture is true,
we designed two variants of the system. In the first variant, users were exposed to just one document from a single
viewpoint at a time, in a fixed order. Users could, however,
explicitly ask for the next document to be one of opposing
viewpoint. If not, the next relevant result was shown by default. UI variants 1a and 1b (cf. Table 1) follow this setting,
and Fig. 1 shows an example interface.
In the second variant, users were exposed to the contrasting evidence right at the start. The primary document and a
document of contrasting viewpoint were shown side-by-side.
Users could still pick which document to read first, and may
choose to ignore the contrasting viewpoint if they wish to.
UI variants 2a and 2b (cf. Table 1) follow this setting, and
Fig. 2 shows an example.
2. Single document per page vs. multiple documents: The next factor we investigated was how the amount
of information on a page affected the reading pattern and
choice of documents. When fewer documents are shown, humans tend to spend more time reading documents that are
shown, before moving on to the next page. We wanted to investigate if this hypothesis was correct. We designed a third
variant where instead of one document, five documents and
their corresponding counter-arguments were shown to the
user on a single page. UI variants {3, 4a, 4b, 5} show multiple documents per page, and Fig. 3 shows an example.
5. SETTING UP THE USER STUDY
Data and Study topics: We enabled the study for two
controversial issues, one from the health domain and the
other from politics. The primary claims (“issues at hand ”)
included in this study were: (i) Milk: Drinking milk is
a healthy choice for humans; and (ii) Energy: Alternate
sources of energy are viable alternatives to fossil fuels. We
will refer to the corresponding study sessions as the Milk
and Energy tasks, respectively.
1907
For both issues, we collected over 350 snippets of text from
ProCon.org1 , a non-partisan, non-profit public charity website. The website offers quotes relevant to sub-topics within
the issue, categorized as pro (in favor of the question being
asked), con (against the question being asked), or neither
pro nor con. The website also gives an expertise rating on a
five-point scale to each source. Our analysis showed that for
both issues, almost all sources were assigned either a 3-star
or a 1-star rating. We used these manually assigned ratings
as our bimodal rating scheme.
We indexed the passages using the Lemur toolkit2 . We
fixed the retrieval query as a weighted combination of key
words relevant to the topics, and retrieved top 50 passages.
This helped us control the exact set and sequence of passages
that subjects would see in the user study.
Inviting subjects: Volunteers were invited to participate in the study by announcing the study on mailing lists in
many departments within a large, diverse, public university.
Invitation emails were sent out primarily to graduate students, staff members, and to participants of a nation-wide
multi-center research meeting. The announcement invited
volunteers to participate in a learning task, where they were
expected to learn about a topic and answer questions. Subjects were not informed about the exact nature of the study
or what factors were being measured. They were, however,
informed that they can participate in the learning tasks related to two topics, that each task would take about 45
minutes to complete, and that they will be rewarded a fixed
amount for each task they successfully complete.
Volunteers who agreed to participate in the study were
issued a unique identifier (pseudonym) that they would use
to access the study. The pseudonyms were statically mapped
to two interface variants from the ones listed in Table 1.
One task was assigned a variant from the first four that
showed one or two documents per page, and the other task
was assigned a variant from the latter four that showed ten
documents per page. Subjects were free to choose any of
the two tasks first, so the researchers did not have control
over which interface variant subjects saw, and for subjects
that took part in both topics, the order in which they saw
those variants. All responses and interactions were recorded
using the pseudonym and no identifiable information was
requested or recorded during the study.
6.
Figure 4: Variation of number of documents read for
each document position. Documents in the primary
set are in odd positions (shaded green) while those in
the contrast set are in even positions (shaded white).
(i) Explicitly showing contrast documents helps
Subjects read an average of 18.6 documents, but considered as many as 31 documents, including ones they choose
to skip. To understand which documents subjects read and
which they skip, we compared how many times subjects read
a document per result position. In UI# 1a and 1b, subjects
are shown only one document by default, from the primary
document set; and they can choose to see the corresponding contrast document next. For all other UI variants, the
contrast documents are shown along side the primary documents, with documents in favor of the sub-topic on the left
column, and documents against the sub-topic on the right.
Fig. 4 shows how the readership changes with document
position for the top 10 results. We compare two scenarios
– one in which only one primary document is shown by default (UI# {1a, 1b}) and the other where one primary and
one contrast document are shown side-by-side (UI# {2a,
2b}). As we see, the readership for contrast documents is
significantly lower when it is not shown by default, while
the readership of the primary document does not change as
much. This shows that users do not tend to pro-actively ask
for the contrasting viewpoint. But when shown side-by-side,
they tend to read both viewpoints.
ANALYSIS OF STUDY RESULTS
We invited 24 volunteers to participate in the user study,
and the average age of participants was 28.6 ± 4.9 years.
Each participant could take part in at most two study tasks.
In all, we collected information from 40 study tasks, with
most participants choosing to take part in both tasks. The
profile of how subjects interacted with the system was similar for both tasks. Typically, subjects took 7–10 minutes
to complete each of the pre-study and post-study questionnaires, and spent about 26.5 minutes in the study phase.
(ii) Showing multiple documents per page helps
When multiple documents are shown per page, users tend
to be more selective in what they read. Table 2 summarizes
the variation in reading pattern as number of documents
are increased. Showing one document per page not only significantly reduce the total number of documents read, but
subjects also spend more time reading those documents onan-average. By showing the contrast document along side,
subjects tend to spend more time overall to read more documents. When multiple documents are shown on a page, subjects are able to consider and skip many more documents in
the stipulated time. When we compare single document and
multi-document views, we see that although subjects read
the same number of documents overall, they scanned about
15 more passages in the multiple document view.
6.1 Which documents do subjects read/skip?
In this paper, we focus on the factors that help users get
a balanced understanding of the topic. Specifically, we want
to understand how contrasting viewpoints and expertise ratings help users decide what documents to read.
1
2
http://www.procon.org/
http://www.lemurproject.org
1908
UI#
#passages
Number of docs read
Number of docs skipped
Study time (in min)
Time/doc read (in min)
{1a, 1b}
1
13.5
4.3
26.3
2.38
{2a, 2b}
2
20.3
5.1
29.4
1.77
{3, 4a, 4b, 5}
10
20.2
19.7
25.4
1.43
learn about a controversial topic. We find that explicitly
showing contrasting viewpoints is significantly better than
merely giving an option to users to look at documents from
alternate viewpoints. Moreover, showing expertise rating
helps users pick which documents to read and which to omit.
We also observe that these factors help users reduce strong
biases and learn new topics in an unbiased fashion [9].
Although this is an initial study on how to present controversial topics to potentially biased users, the findings are
already interesting. Subjects spent over an hour for each
task and gave valuable feedback during the course of their
participation. The insights gained by this study would help
us and other researchers optimize the design of an automated claim verification system, that not only learns which
evidence documents are most relevant to show to the user,
but also what additional information needs to be provided
to help users assimilate the information faster.
This study can be further extended to understand how
human biases affect the perceived quality of the documents.
Further, based on users’ prior knowledge, a different set of
results may be presented. A larger scale study is planned
to look into these aspects. Some subjects suggested that
an overall summary of relevant sub-topics would help them
understand the issues better. This is indeed an interesting
need, but beyond the scope of the current study.
Table 2: Interface type influences reading pattern.
Scheme
1⋆
2⋆
3⋆
4⋆
5⋆
Single document view (UI# {1a, 1b, 2a, 2b})
Bimodal
76.3%
85.4%
Uniform
56.3% 77.3% 85.3% 83.2% 84.3%
Multiple document view (UI# {3, 4a, 4b})
Bimodal
39.7%
61.3%
Uniform
38.2% 52.8% 60.7% 66.0% 66.9%
Table 3: Readership increases with expertise rating.
(iii) Higher expertise rating gets higher readership
Next, we looked at the impact of expertise rating on which
documents are read and which are skipped. For this analysis,
we distinguish the UI variants based on whether the rating
scheme was bimodal or random. Table 3 shows the variation
in documents read for these two rating schemes.
We focus on two classes of interfaces – one in which a
single primary document (and the corresponding contrast
document) was shown, and the other where five primary
documents were shown. In both cases, we find that the
documents that are rated high are read more often than
those with poor expertise ratings. This is especially true
when multiple documents are shown and subjects have more
choice in what to read. As we see in Table 3 (last row), only
38% of documents with a 1-star rating were read, whereas
about 67% of documents with a 5-star rating were read by
the subjects. When the documents are rated under the bimodal scheme, the documents rated high are 22% more likely
to be read than the documents rated low. These trends also
follow when only one document (with or without the contrast document) is shown to the subjects. We also see that,
in this scheme, relatively fewer documents are skipped. This
is because, in the single document view, there is typically
just one option – i.e. to either read or skip the document,
without the knowledge of what the next document would
be. So, a relatively larger number of documents are read,
even if they have low expertise ratings.
8. ACKNOWLEDGMENTS
This research is partially funded by grants from the MIAS
Center of Excellence of the Department of Homeland Security, the Information Trust Institute, the Army Research
Laboratory, and the Office of Naval Research.
9. REFERENCES
[1] J. Baron. Thinking and Deciding. Cambridge Press,
2000.
[2] A. Galland, S. Abiteboul, A. Marian, and P. Senellart.
Corroborating Information from Disagreeing Views. In
Proc. of WSDM, pages 131–140, 2010.
[3] E. Pariser. The Filter Bubble: What the Internet is
Hiding from You. Penguin Press, 2011.
[4] P. Pirolli, E. Wollny, and B. Suh. So You Know You’re
Getting the Best Possible Information: A Tool that
Increases Wikipedia Credibility. In Proc. of CHI,
pages 1505–1508, 2009.
[5] S. Plous. The Psychology of Judgment and Decision
Making. McGraw-Hill, 1993.
[6] B. Suh, E. H. Chi, A. Kittur, and B. A. Pendleton.
Lifting the Veil: Improving Accountability and Social
Transparency in Wikipedia with WikiDashboard. In
Proc. of CHI, pages 1037–1040, 2008.
[7] C. S. Taber and M. Lodge. Motivated Skepticism in
the Evaluation of Political Beliefs. American Journal
of Political Science, 50(3):755–769, July 2006.
[8] J. Ugander, L. Backstrom, C. Marlow, and
J. Kleinberg. Structural Diversity in Social Contagion.
In Proceedings of the National Academy of Sciences of
the United States of America, pages 1–5, 2012.
[9] V. Vydiswaran, C. Zhai, D. Roth, and P. Pirolli.
Unbiased Learning of Controversial Topics. In Proc. of
ASIST, 2012.
[10] M. Wu and A. Marian. Corroborating Answers from
Multiple Web Sources. In WebDB, pages 1–6, 2007.
(iv) Absence of rating helps low quality documents
In one of the UI variants (UI# 5), all expertise ratings were
hidden from the subjects. We find that, under this setting,
49.8% of the shown documents were read by the subjects.
Comparing to Table 3, we find that unrated documents are
more likely to be read than those with 1-star rating.
7.
CONCLUSION AND FUTURE WORK
Providing access to unbiased and credible information is
critical to satisfy information need in many domains. Moreover, when verifying controversial claims, it is also important
to understand and help users overcome the tendency to stick
to one’s own viewpoints. We conducted a user study called
BiasTrust to understand what factors significantly help users
1909